Refactor/extra engine types

[johnretsas]

Parser detection driven by the type (types.ts + ast-parser.ts)

The AST parser's list of recognized step.* methods was maintained by hand,
disconnected from the actual step API (StepBaseContext). That let the two drift,
and the name→StepType mapping relied on an unchecked as StepType cast plus a
special-case sleep branch.

• Added STEP_BASE_METHOD_TYPES, derived from
  _STEP_BASE_METHOD_TO_TYPE: Record<keyof StepBaseContext, StepType>. The
  Record<keyof StepBaseContext, …> annotation makes the compiler enforce two
  things: every step method is mapped (add a method to StepBaseContext →
  compile error until it's in the map), and every value is a real StepType.
• The parser's || chain of method-name comparisons + as StepType cast +
  sleep-alias ternary collapse into a single STEP_BASE_METHOD_TYPES.get(methodName)
  lookup. undefined doubles as the "not a step method" check, and the
  sleep → DELAY alias is now data in the map rather than a branch in the parser.

Engine type-safety & structure (engine.ts)

• Typed StepTypeToIcon as Record<StepType, string> so a new StepType forces
  a matching icon entry.
• Removed a redundant as WorkflowInternalDefinition cast on the
  workflows.set(...) literal. The spread already produces the correct type, and
  the cast was suppressing excess-property checking (stray keys are now caught).
• Extracted the ~110-line baseStep object literal out of handleWorkflowRun
  into a private createBaseStep(run, workflowId, runId) method.
• Added a private createChildWorkflowRun(...) that wraps createWorkflowRun
  with parentRunId/parentStepId required at the type level, so the
  child-invocation path can't be wired up without parent linkage.

Fail loudly on a foreign workflow_runs table (migration.ts)

Previously, if a consumer already had a workflow_runs table in the target
schema, migrations could ALTER/index a table pg-workflows didn't create. The
migration now checks (only on a fresh install, i.e. no workflow_schema_version)
whether a workflow_runs table already exists and throws a clear error
pointing the user at a dedicated schema/rename — before any DDL runs.

This is the one behavioral change in the PR; it only fires for genuine collisions
and is otherwise inert (healthy installs are already past version 0).

Test coverage

• Parser (ast-parser.test.ts): added real expect(...).toEqual(...)
  assertions to the previously no-assert switch, nested conditional/loop, and
  mixed-step tests; added a test documenting the known limitation that
  externally-defined handlers aren't parsed (yields [] steps).
• Migration (migration.test.ts, new): fresh DB migrates to the current
  version; migrations are idempotent across repeated runs; a foreign
  workflow_runs table makes runMigrations throw without touching it.
• Idempotency (engine.test.ts): reusing an idempotencyKey enqueues no
  second job and leaves the first run untouched (completed case); and the same
  holds while the first run is still mid-execution (paused) — the duplicate is a
  no-op and the in-flight run is undisturbed.

Testing

• tsc --noEmit clean
• Full unit suite passes (parser, engine, definition, migration)